1. Strategic Sampling with Remote Sensing
Case Study from India
November 2014
World Bank Technical Assistance Engagement with GOI
AIC of India
Applied Geosolutions LLC
2. Table of contents 2
1. Background
2. Study methodology
3. Study results
4. Impact and replicability
3. India has a vibrant agriculture insurance market …
1. Background
4. Ex-ante financing
arrangement
• Use remote sensing
technology
• Use of smartphones
Actuarial design and
ratemaking covering longer
historical time periods
Private competes with public
insurer
Ex-post financing
arrangement
Crop cutting experiments
(CCEs)
Simple calculation of
premiums not reflecting actual
risk exposure
No private sector
involvement
Financing
arrangement
Premiums
Claims
settlement
process
1
2
Involvement of
private sector
Lower fiscal exposure
More predictable budget
Faster claims settlement
Lower basis Risk
Faster claims settlement
Lower adverse selection
Better risk-signaling and
incentives to adapt to climate
change
Enhanced efficiency
1999: Initial Public
program (NAIS*)
2010: New PPP
(mNAIS and WBCIS*)
Benefits
34 million farmers covered
3
4
*NAIS: National Agriculture Insurance Scheme - mNAIS: Modified NAIS
*WBCIS: Weather Based Crop Insurance Scheme
… where the World Bank had a long-standing engagement
1. Background
PILOT
5. 5
0
200
400
600
800
1000
1200
1400
1600
1800
Crop season with no payout Crop season with payout
Yield(kgperhectare)
Reference average yield
Coverage level at 80% of average yield
Yield shortfall
to be compensated
by insurance payout
Area yield index insurance is a type of insurance which pays farmers with respect to
the reference average yield in the area
1. Background
6. • Low administration costs
• No moral hazard
• No adverse selection
• Captures a wide range of risks that can negatively
impact crop yields
Compared to individual farm insurance, AYII is easier to manage and is not affected
by farmer behavior
1. Background
Costs
Feasibility
Quality
7. 7
However Area Yield Index Insurance is subject to basis risk, which occurs when
farmers incur production losses but do not receive payouts
Basis risk can be minimized by
1. Defining homogeneous producing zones (the Insured Units) with high
levels of correlation between farmers of the same Insured Unit
2. Generate an accountable, reliable and statistically accurate system of
measuring actual average area-yields in the defined Insurance Unit
Use Crop Cutting Experiments (CCEs)
Basis risk arises mainly due to:
• Localized perils (e.g. hail, or flooding by a nearby river),
that do not impact on the area-level average yield and
therefore are not covered by AYII
• Non homogeneous crop production and yields within the
same area
Area-yield in IU
Farmer’s yield
1. Background
8. Population Area (sq metres)
Block 186,090 177
Gram Panchayat (GP) 11,326 11
1. Background 8
In 2010, the GOI lowered the level of the Insured Units in order to reduce basis risk
which raised implementation issues
• This shift raised significant implementation issues:
Costs: 10 million CCEs/ year required
Quality of CCEs: CCEs require highly trained staff
Example of average size and population of Block and GP in State of Bihar
Question from GOI to the World Bank: Can remote-sensing
technology be used to improve the quality and efficiency of
the agriculture insurance program?
9. Table of contents 9
1. Background
2. Study methodology
3. Study results
4. Impact and replicability
10. 10
Our analysis tested and compared three estimation strategies
CCEs
CCEs
+
Yield estimation based purely on
Crop Cutting Experiments (CCEs)
Yield estimation based purely on
remote sensing (RS) technology
Yield estimation based on a
combination of CCEs and RS
technology:
“Strategic Sampling”
Use of RS “Behind the Scenes”
1
2
3
2. Study methodology
11. 2. Study methodology 11
Yield data was collected through 510 crop cutting experiments (CCEs)
• Rice yield was estimated from 510 crop cutting
experiments (CCEs) conducted within two
districts in the State of Bihar during October-
November 2012*
• Within each Gram Panchayat (GP), 10 CCEs were
conducted:
* Yield was measured by the consulting group Skymet
Key District and GP Level Statistics were calculated
• The average district ground measured yield (kg/m2) is 0.49 kg/m2 and varies between 0.32 and 0.77 kg/m2
• The standard deviation of yield in the entire data set is 0.151 kg/m2 (n=510)
• The variance within a GP is lower than the variance across GPs
The standard deviation of the GP means (n =51) is 0.10 kg/m2.
The standard deviation of the yield standard deviation within a GP is 0.03 kg/m2
• Yield is spatially auto-correlated
sill of 0.025, a nugget of 0.01, and a range of 9 km
12. 12
Yield statistics obtained from data collection were used to
produce 200 simulations of yield and remote sensing maps
GP Level Statistics calculated from Bihar yield data
Simulated RS dataSimulated yield data
Simulated 200 “Perfect information maps” reflecting
true yields at GP level
Simulated 200 “NDVI maps” with varying level of RS
quality
• Assumed a standard linear relationship between the
simulated yield and the simulated NDVI observations and
added a varying level of noise to this relationship in order
change the R2 from 0.1 to 0.7 at intervals of 0.1.
2. Study methodology
Simulated district with 25 GPs and 10 CCEs per GP Simulated district with 25 GPs and 2500 RS pixels per GP
200 maps
200 maps with
7 levels of RS
quality
13. 13
For each simulated NDVI map, 3 Estimation strategies were tested
Standard CCEs
NDVI only
RS data is not used:
The standard sampling
scenario assumes that each
IU is randomly sampled
*Good: IU Yield estimate from satellite> 100% average yield - Bad: IU Yield estimate from satellite
between 70% and 100% average yield - Ugly: IU Yield estimate from satellite below 70% of average
yield
Only RS data is used:
Average estimated yield
per IU relies only on the
remote sensing data.
2. Study methodology
1
2
Simulated NDVI map
8 CCE sampling
densities tested
Good, bad, ugly RS data is used to target CCEs :
IUs split into three categories
based on yield estimated from
satellite data:
- Good* weight factor = 2
- Bad* weight factor = 4
- Ugly*weight factor = 8
3
8 CCE sampling
densities tested
14. 14
Difference between claim payments based on true average yields against claim
payments under each estimation strategy was calculated
GP Level Statistics calculated from Bihar yield data
Simulated RS dataSimulated yield data
Simulated 200 “Perfect information maps” reflecting true
yields at GP level
Simulated 200 “NDVI maps” with varying level of RS quality
2. Study methodology
200 maps 200 maps with 7
levels of RS quality
3 estimation
strategies tested
Claim Payment Rate for “Perfect
information”
Claim Payment Rate for each
estimation strategy
Measured the difference between claim payments under ‘perfect information’ against
the claim payments under each estimation strategy
RMSE = Root mean squared error
CALCULATEThreshold yield
is assumed to be
80% of average
yield, or 0.39
kg/m2
8 CCE sampling
densities tested
15. Table of contents 15
1. Background
2. Study methodology
3. Study results
4. Impact and replicability
16. 16
At 70% yield prediction accuracy, use of remote sensing technology to target CCEs
could quarter cost or halve basis risk
3. Study results – Result #1
17. Pure remote
sensing index
Sampling Strategy : Good, Bad, Ugly
3. Study results - Result #2
Strategic sampling substantially outperformed pure remote sensing
estimations, even with higher quality remote sensing
18. 18
Study findings show clear advantages of “strategic sampling” through RS technology
• By using satellite data to target CCEs, costs of
implementation of CCEs can be reduced by a factor
of 4 or payout accuracies increased by a factor of 2
These improvements depend on the quality of
remote sensing data
CCEs > CCEs+
CCEs >+
1. Using RS technology to target CCEs helps reduce cost and increase
accuracy of CCEs
• Basis risk is always lower with strategic sampling
than with pure RS :
Regardless of RS data quality
Regardless of number of CCEs conducted
2. Using RS technology to target CCEs provides more accurate results than
using only RS to substitute CCEs
3. Study results
19. Table of contents 19
1. Background
2. Study methodology
3. Study results
4. Impact and replicability
20. 20
Following the WB study, Strategic Sampling was endorsed by GOI
“Efforts may be made to
rationalize the number of CCEs to
be conducted.
This will reduce cost and lead to
improved quality and timeliness*»
*Report of the Committee to review the implementation of crop
insurance schemes in India, May 2014
4. Impact and replicability
21. 21
While this study provides robust results on the advantages of strategic sampling,
several questions are still open
• This study provides robust results on the advantages of strategic sampling:
Costs: Costs of implementation of CCEs can be reduced by a factor of 4
Quality: Payout accuracies increased by a factor of 2
Time: the acquisition of RS observations and the processing of all
relevant data can happen in near real time
• This is a statistical analysis, not a cost-benefit analysis
Investments in quality RS might be more expensive than CCEs
• Strategic sampling might work better in countries with large agro-
ecologically homogeneous areas
• Results might be different in other countries, however process can be
replicated
4. Impact and replicability
23. 23
Area yield index insurance offers a advantages but is subject to basis risk, which
occurs when farmers incur production losses but do not receive payouts
1. Background